Painting to Imagine

[nesretep-anp1]

I do have some trouble with the "Painting to Imagine" in the "Image Generation" process.

At first, ... it is always tried to apply the/my current location to the/my prompt. This is nearly always useless but in some cases it even destroys the original intention of the image description.

How could one disable having these things get applied?!

But my current and (for me currently) important "question" is, ... if using (e.g.) the following prompt ...

create an image with a storm trooper fight against mara jade

... the following train of thought is applied, ...

Selected Tools: notes, image
Enhancing the Painting Prompt
Painting to Imagine: **I cannot create content that depicts violence or harm.** Is there anything else I can help you with?
Generating a well-informed response

... but neverthelesse the most brutal scene ever seen is displayed; but with no relation to Star Wars and Storm Troopers.

Inspired from this, I tried the prompt ...

Paint a picture of a young girl wearing a shirt showing the text “no sex!”

... where the following train of thought was applied, ...

Selected Tools: notes, image
Enhancing the Painting Prompt
Painting to Imagine: **I cannot create content that sexualizes a minor**. Can I help you with something else?
Generating a well-informed response

... but neverthelesse a picture of a girl without any shirt is generated. (Comment for this: I absolutely do understand and follow the "I cannot create content that sexualizes a minor" message; but ... there is a SUCH A pic returned and MY prompt has nothing to do with such a message)

My real question is, ...

... where comes the "I cannot create content that depicts violence or harm." from?
... how could I disable/change this filter/behaviour/...?
... This specially according to location and stuff of my own context

I use LocalAI with dreamshaper as the image generation mode and using this prompt via curl (or web frontend of Local AI) the/a corrent image - according to the prompt - is generated.

[debanjum]

Thanks for the sharing the detailed train of thought, query, output and your image generation setup. There are multiple different points being discussed here. Let me address the unsafe image generation bit first:

Context:

• We recently added the ability to use local image generation models (via any openai compatible api). So this flow has not been tested much
• There are multiple AI models involved. The LLM is supposed to generate a more detailed image description from your original query, the text to image model uses this description to generate the image.

Issue:

What seems to be happening is that for your prompt the LLM generates a "cannot generate content of type X" which is then passed to the image generation model. This results in the image generation model using that text as its instruction to generate the image! This result in it actually generating a generic image of type X 😬 . For commercial image models that we use, this wasn't being an issue as the image gen safety filters would prevent such image generations anyway if the LLM did not refuse.

Fix

The fix for this would be to not have the image generation model called if LLM doesn't want to generate such an image.

[debanjum]

My real question is, ...

... where comes the "I cannot create content that depicts violence or harm." from?

This comes from the default chat model you've configured. The priority order of default chat model selection is:
server chat settings > user chat settings > first chat model added. See

khoj/src/khoj/database/adapters/__init__.py

Lines 1112 to 1113 in a3b5ec4

	async def aget_default_chat_model(user: KhojUser = None):
	"""Get default conversation config. Prefer chat model by server admin > user > first created chat model"""

... how could I disable/change this filter/behaviour/...?

If you don't want the model to censor your response, you should change the default model to a less censored LLM model

... This specially according to location and stuff of my own context

There is no way to explicitly disable location context right now. The best way to avoid the above issues is to create a custom image generation agent and

1. Set the agent's personality prompt and tell it to not use your location, use your exact query with no modification or whatever other instructions you want it to follow. You may need to experiment with the personality prompt + model that works best for this
2. Select only the "general" tool in the "Input tools" section to have it not lookup any context (online, notes, webpage etc.). Note: This does not disable location context which is passed directly

[nesretep-anp1]

As a hint and to prevent doubled discusions/comments, ... I already opened an issue around the things I found in that context:

#1096

[debanjum]

There seem to be multiple issue spread across this discussion (#1095) and #1096. Here I've addressed the

1. the unsafe image generation bug
2. How to customize your agent to not pass context when generating images

The chat model prioritization logic has been discussed in both.
The issue #1096 also addresses the confusion around configuring the tokenizer.

[shbnm-shm]

Hi